Inhalt

•  
   Contents
•  
   Introduction
•  
   1 Floating-point numbers, standards, and the user environment
•  
   1.1 The foundation: Floating-point numbers
•  
  1.2 The standards: IEEE 754 and 854
•  
   1.2.1 IEEE 754
•  
   1.2.2 IEEE 854
•  
   1.2.3 IEEE 754-2008
•  
   1.2.4 IEEE 754-2018
•  
   1.3 Rounding and error analysis
•  
   1.3.1 Some basic properties
•  
   1.3.2 Stability
•  
   1.3.3 Use cases for rounding
•  
   1.4 The user environment
•  
   1.4.1 Limitations of the user environment definition
•  
   1.4.2 Floating-point hardware
•  
   1.4.3 The operating system
•  
   1.4.4 Floating-point libraries
•  
   1.4.5 Programming languages
•  
   1.4.6 The compiler's role
•  
   1.4.7 Interpreters
•  
   1.4.8 In-network floating-point computations
•  
   1.4.9 Resilience
•  
   1.4.10 Comparing and testing floating-point environments
•  
  2 IeeeCC754
•  
   2.1 History
•  
   2.2 IeeeCC754
•  
  2.3 Testsets
•  
   2.3.1 Addition and subtraction
•  
   2.3.2 Multiplication
•  
   2.3.3 Division
•  
   2.3.4 Square root
•  
   2.3.5 Remainder
•  
   2.3.6 A note on conversions
•  
   2.3.7 Conversions between floating-point formats
•  
   2.3.8 Rounding floating-point numbers to integral values
•  
   2.3.9 Conversion between floating-point and integer formats
•  
   2.3.10 Decimal to binary and binary to decimal conversion
•  
  2.4 Results
•  
   2.4.1 Intel and AMD
•  
   2.4.2 SUN Sparc
•  
   2.4.3 FMLib
•  
   2.4.4 MpIeee
•  
   3 IeeeCC754++
•  
  3.1 IeeeCC754++: Introducing extensions
•  
   3.1.1 Testing IEEE-conformity
•  
   3.1.2 Testing the user environment: Default mode
•  
   3.1.3 Testing parts of a floating-point environment
•  
   3.1.4 Testing distributed floating-point environments
•  
   3.1.5 Supporting arbitrary floating-point environments
•  
   3.1.6 Building for arbitrary environments
•  
   3.1.7 Large-scale testing and analysis
•  
   3.1.8 Studying the influence of compiler options
•  
   3.1.9 Testing modes
•  
   3.1.10 Input and output
•  
   3.1.11 Support for IEEE 754-2008
•  
   3.1.12 Analysis capabilities
•  
   3.1.13 Preserving backwards compatibility
•  
   3.1.14 Technical changes
•  
  3.2 Input and output
•  
   3.2.1 Test vector formats
•  
   3.2.2 Output formats
•  
   3.3 Testing modes
•  
   3.3.1 Classic mode
•  
   3.3.2 Verbose mode
•  
   3.3.3 Default mode
•  
   3.3.4 Distributed computing modes
•  
   3.3.5 Miscellaneous modes
•  
   3.3.6 Common command line options
•  
   3.4 The evaluation framework
•  
   3.4.1 Using the evaluation framework
•  
   3.4.2 Analysis modules
•  
   3.4.3 Tools for (semi-)automated testing
•  
   3.5 The optimisation framework
•  
   3.5.1 Using the optimisation framework
•  
   3.5.2 Fitness modules and adding fitness functions
•  
   3.5.3 Timing the effect of compiler options
•  
  4 Extended testsets
•  
   4.1 General considerations
•  
   4.1.1 The Table Maker's Dilemma
•  
   4.1.2 Adding operators and test vectors
•  
  4.2 Fused multiply-add
•  
   4.2.1 Testset
•  
   4.2.2 Considerations concerning portability
•  
   4.3 Powers and roots
•  
   4.3.1 Testsets
•  
   4.4 Trigonometric functions
•  
   4.4.1 Testsets
•  
  4.5 Exponentials and logarithms
•  
   4.5.1 Testsets
•  
  4.6 Miscellaneous functions
•  
   4.6.1 Testsets
•  
   4.7 Generating testsets
•  
   4.7.1 A note on precisions
•  
   4.7.2 convertTestsets.py
•  
   4.7.3 genUCB.sh
•  
   5 Architecture ports
•  
  5.1 The default architecture
•  
   5.1.1 The default port
•  
   5.1.2 The dummy port
•  
   5.2 The x86 architecture
•  
   5.2.1 The x86 port
•  
   5.2.2 The mic port
•  
   5.3 The ARM architecture
•  
   5.3.1 The arm port
•  
   5.3.2 The aarch64 port
•  
   5.4 The Power Architecture
•  
   5.4.1 The ppc port
•  
   5.4.2 The cell port
•  
   5.4.3 The bgq port
•  
   5.5 GPUs and accelerators
•  
   5.5.1 The nv port
•  
   5.5.2 The opencl port
•  
  5.6 In-network computations
•  
   5.6.1 The mpi port
•  
  5.7 Virtual machines and software libraries
•  
   5.7.1 The java port
•  
   5.7.2 The softfloat port
•  
   5.7.3 The mpfr port
•  
   5.7.4 The crlibm port
•  
   6 Selected results
•  
  6.1 A detailed example
•  
   6.1.1 User environments
•  
   6.1.2 Manual testing procedure
•  
   6.1.3 (Semi-)Automated testing procedure
•  
   6.1.4 Analysing the logfiles
•  
   6.2 Different compilers
•  
  6.3 x86
•  
   6.3.1 Xeon
•  
   6.3.2 Xeon Phi: KNL
•  
  6.4 ARM
•  
   6.4.1 ARM: VFP, NEON
•  
   6.4.2 AARCH64: ASIMD, SVE
•  
   6.5 PowerPC
•  
   6.5.1 POWER8
•  
   6.5.2 Cell
•  
  6.6 Accelerators
•  
   6.6.1 CUDA
•  
   6.6.2 OpenCL
•  
   6.7 Software
•  
   6.7.1 SoftFloat
•  
   6.7.2 MPFR
•  
   6.7.3 Java
•  
   6.8 Elementary functions
•  
   6.8.1 C99 vs. C++11
•  
   6.8.2 Trigonometric operators
•  
   6.8.3 Exponentials and logarithms
•  
   6.8.4 Power operators
•  
   6.8.5 roundTiesToEven results
•  
   6.9 Optimisation framework
•  
   6.9.1 User environments
•  
   6.9.2 Two-step process
•  
   6.9.3 Example run with [style=mhplain]sixloops
•  
   6.9.4 Example run with HPCG
•  
   6.10 Result summary
•  
   6.10.1 Basic operations and conversions
•  
   6.10.2 Elementary operators
•  
   6.10.3 Some notes on applications
•  
   Summary and outlook
•  
  A The IeeeCC754++ build system
•  
   A.1 Changes to the code base
•  
   A.1.1 IeeeCC754 code structure
•  
   A.1.2 IeeeCC754++ code structure
•  
   A.2 The build system
•  
  A.3 Configuring and building IeeeCC754++
•  
   A.3.1 Building overview
•  
   A.3.2 Choosing an architecture
•  
   A.3.3 Choosing FPUs
•  
   A.3.4 Choosing the compiler and compiler options
•  
   A.3.5 Generic build features
•  
   A.3.6 Additional build options
•  
   A.3.7 Cross compilation
•  
   A.3.8 Building historic modes
•  
   A.3.9 A detailed example
•  
  B Adding a new architecture to IeeeCC754++
•  
   B.1 File structure
•  
   B.2 Build system: configure.ac
•  
   B.3 Build system: Makefile.am
•  
   B.4 Implementing the new architecture
•  
   B.4.1 src/xyz/DriverFloat_main.h
•  
   B.4.2 src/xyz/fpu_main.cc
•  
   B.5 Adding an FPU
•  
   B.6 Implementing an operation
•  
   B.7 Handling vector FPUs
•  
  B.8 Initialising an FPU
•  
   B.8.1 Registering operations and rounding modes
•  
   B.8.2 Enabling vector FPUs
•  
   B.9 Example code for the new architecture and FPU
•  
   B.10 Setting up main()
•  
   B.11 Setting up and building the new architecture
•  
  C Reference material
•  
   C.1 fma example
•  
   C.2 IeeeCC754++ usage
•  
   C.3 IeeeCC754++ classic mode usage
•  
   C.4 Error codes used in IeeeCC754++
•  
   C.5 Reference task files
•  
   C.6 Evaluation function example
•  
   C.7 Fitness function example
•  
   List of Figures
•  
   List of Tables
•  
   List of Listings
•  
   Bibliography